David Ho is about to enter his fifth year in Purdue University’s Ph.D. program in electrical and computer engineering where he specializes in image processing and computer vision research. Ho moved to the US from Gwangju, Korea during high school, and then attended the University of Illinois at Urbana-Champaign to study for both undergraduate and Masters degrees in electrical and computer engineering. This summer, Ho has been working on a collaboration between HP’s Print Software Platform organization and Emerging Compute Lab, called Pixel Intelligence, applying his expertise in image segmentation to the challenge of picking out people in any specific image.
HP: Can you tell us more about your internship project?
I’ve been using deep learning to improve what we call person segmentation, which is where a computer is able to separate the image of a person from any background. Humans can distinguish between different kinds of images very easily. But computers just see images as an array of pixel values. So we need to find ways to make computers “understand” images of people as people.
HP: How have you been doing that?
I’ve been taking several existing data sets of images where we have already established the “ground truth” of the images and using those data sets to teach a computer program what a person looks like. Once it is trained, I input new images and see how well the program can pick people out of them. The idea is to reduce the number of errors we get in doing that, and to be able to do it faster.
HP: How has it been going?
We’ve had some good results. One thing we’ve been able to do is get this running on a webcam camera, so that it can segment out people in every frame it records.
HP: What’s the challenge in doing that?
One is getting it to work for a relatively crude camera. Another, which we’re still working on, is reducing the processing required to do the segmentation. So far we’ve been running it on a processing unit designed for heavy computation. But we’d like to be able to run it on a smaller device.
HP: Will this work feature in your Ph.D. thesis?
Not directly. In my Ph.D., I’m also looking at applying deep learning to image processing, but I’m looking at understanding microscope images and segmenting out different biological structures. So the application is different but the main idea is the same: helping computers to make sense of interesting images.
HP: Is this your first time interning at HP Labs?
Yes, and it’s my first internship in an industrial lab.
HP: What has struck you as different about working in an industrial lab setting?
I’ve been impressed how industrial labs value creating software that anyone can use. My segmentation solution was pretty good, for example, but required a lot of processing power. So my mentor, Dr. Qian Lin, has pushed me to make it smaller so it’s of more value to more people.